[ Mechanical Translation , vol.3, no.2, November 1956; pp. 52-61] Mechanical Translation of French † L. Brandwood, Birkbeck College Research Laboratory, University of London IN THE NEAR FUTURE an attempt to translate from a foreign language by machine will be made at the computational laboratory of Birk- beck College. It will differ from previous ex- periments in that the sentences or passages of French to be translated will not have been spe- cially chosen or "doctored" in any way before- hand: on the contrary, they will be constructed by French scholars, who will be invited to do their best to fault the machine. What follows is an account of the method, or program, which makes such an experiment possible. First of all, however, I think a brief explana- tion of the general procedure in handling words by electronic computer will not be out of place. To a layman like myself the information that such a computer works in terms of electronic impulses conveys very little, but as I suspect that the majority of readers are in the same boat, I shall no doubt be forgiven for using terms which though strictly not correct have at least the advantage of being more readily intelligible. We may say then that the computer works, like any other calculating machine, with digits. It cannot recognize letters; at any rate not as they stand. In order that it may recognize them the alphabet has first to be coded, that is to say, each letter represented by a particular number. Since a word is merely a collection of letters and we have a number for each letter, we can now put any words we choose into the computer to be stored on a magnetic drum in terms of impulses. These constitute the ma- chine's "dictionary", which may be looked upon as exactly like an ordinary French- English dictionary, but with the numbers repre- senting the French words arranged in order of either ascending or descending magnitude. This is important for the rapid identification of the words in the text being translated, which takes place as follows. The machine identifies an incoming word by comparing the number which represents it with the numbers contained † An earlier version of this paper appeared in Babel, Vol.11, No. 3, 1956, a special issue devoted to mechanical translation. in its dictionary. The comparison takes the form of a simple subtraction. Starting with a number at the middle of the dictionary it sub- tracts from this the number of the word being translated: if the result is zero, the word has been identified and the English equivalent, which is marked alongside the French can be output; if the result is not zero, then the num- ber sought lies in one or other of the two halves of the thus divided dictionary according to whether the result is positive or negative. So, taking the number at the middle of the relevant half, the machine carries out the same compa- rison as before, and if as is more than likely this too proves unsuccessful, it con- tinues halving the numbers that remain until the one required is located. This ingenious process of elimination, which was devised by Dr. A.D. Booth, ensures the speedy identification of words even in a very large dictionary. In fact the size of the latter can be greatly increased without appreciably lengthening the look-up process. To be pre- cise, the number of comparisons required is doubled as the number of words in the diction- ary is squared. This means that since on an average seven comparisons are required to identify any word in a dictionary of a hundred, only fourteen will be necessary to cope with a dictionary of ten thousand or twenty-eight to cope with one of a hundred million. The capabilities of the look-up method might seem rather wasted when it is revealed that the maximum storage capacity of the Birkbeck computer is at present only 250 dictionary items (i.e. 250 foreign words each with one English equivalent), though this will shortly be increased to 2000. It must be borne in mind, however, that this is not the limit of all computers, even of those existing, and storage capacities will continue to grow larger. But that lies in the future, and for the time being all our attempts at translation have to be scaled down and done, so to speak, in miniature. The machine's capacity allowed us 250 dic- tionary items. We could not use all of these for the actual dictionary, however, because we wished to include a few elementary syntactical instructions to enable the machine to produce Mechanical Translation of French 53 something a little better than a word-for-word translation. We did not know beforehand in what proportion the dictionary space should be divided between vocabulary and grammar, but reckoned that with a vocabulary of about a hundred words there would be enough space left for the instructions in mind and a little over for subsequent modifications. What was wanted therefore, was a vocabulary which, though very small, would yet be capable of producing good French sentences in a reasonable variety; in short, the proportion of the various parts of speech which it contained had to be correctly balanced. This was achieved in a very simple fashion by selecting a passage from a book and extracting the first hundred different words: to these were added any missing forms of the personal, possessive and relative pronouns en- countered. The result was a vocabulary con- sisting of 25 verbs, 2 adverbs, 23 nouns, 10 adjectives, 3 numerals, 8 prepositions, 6 con- junctions, all forms of the unstressed personal pronoun (except en and y), all forms of the simple possessive pronoun, the relatives qui, que, dont, the definite and indefinite articles and the negative ne. . . .pas. Volunteers will be presented with this vocabulary and asked to con- struct sentences or short passages for the ma- chine to translate. They will be permitted to use the singular and plural forms of all nouns and adjectives, the masculine and feminine of the adjectives, and all parts of the verb except the 1st person imperative and the subjunctive mood. The former is excluded because it re- quires a rather periphrastic translation, which combined with its comparative rarity makes it hardly worth bothering about, the latter because many of its forms are identical with those of the present and imperfect indicative, and could only be distinguished from these by taking into account the syntactical context. This, however, is out of the question with the present storage capacity. Nevertheless, dis- regarding the compound tenses, which are ac- counted for by the separate consideration of avoir or être and the perfect participle of each verb, there still remain about 30 different forms to each regular verb, the actual figures for the four conjugations being: donner 33, finir 29, vendre 32, recevoir 31. For an irregular verb, unless it is defective, the number is usually a little higher. In short, the 25 verbs (13 regu- lar, 12 irregular) produce well over 800 differ- ent forms, all of which the machine must be able to translate. Obviously, such a large number will not fit into our machine's dictionary as they stand. How then are they to be incor- porated? The solution thought of as long ago as 1947 by Booth and Richens, is that a word should be divided into a stem and an ending. Here we must pause in order to define our terms. A stem is not necessarily a stem in the sense that grammarians know it. Generally speaking, it is the longest part of a word common to all forms (inflections) of that word. The stem of donner, for example, is donn-; what remains of each inflection of donner after this has been subtracted constitutes an ending. This, however, is not a universal rule, since there are instances where a single stem would account for all forms of the verb, but more than one is chosen for the sake of utility (e. g. the stem acqu- would cover all forms of acquérir, but in addition acquier- has to be used in order to comply with the general system covering all verbs). Utility might in fact be said to be the only principle of grammar in me- chanical translation. Verbs are divided, then, into stems and end- ings, and the following results cover each of the Four Regular Conjugations: I. Donner (stem: donn-) Person Singular Plural Tense 1st 2nd 3rd 1st 2nd 3rd Present e es e ons ez ent Imperative e e ez ent Future rai ras ra rons rez ront Conditional rais rais rait rions riez raient Imperfect ais ais ait ions iez aient Past Hist. ai as a âmes âtes erent Infinitive -er Present Participle ant, ante, ants, antes Perfect Participle é, ée, és, ées 54 L. Brandwood II. Finir (stem: fin-) Person Singular Plural Tense 1st 2nd 3rd 1st 2nd 3rd Present is is it issons issez issent Imperative is isse issez issent Future irai iras ira irons irez iront Conditional irais irais irait irions iriez iraient Imperfect issais issais issait issions issiez issaient Past Hist. is is it îmes îtes irent Infinitive -ir Present Participle issant, issante, issants, issantes Perfect Participle i, ie, is, ies III. Recevoir (stem: rec-) † Person Singular Plural Tense 1st 2nd 3rd 1st 2nd 3rd Present ois ois oit evons evez oivent Imperative ois oive evez oivent Future evrai evras evra evrons evrez evront Conditional evrais evrais evrait evrions evriez evraient Imperfect evais evais evait evions eviez evaient Past. Hist. us us ut û mes ûtes urent Infinitive -evoir Present Participle evant, evante, evants, evantes Perfect Participle u, ue, us, ues IV. Vendre (stem: vend-) Person Singular Plural Tense 1st 2nd 3rd 1st 2nd 3rd Present s s - ons ez ent Imperative s e ez ent Future rai ras ra rons rez ront Conditional rais rais rait rions riez raient Imperfect ais ais ait ions iez aient Past. Hist. is is it îmes îtes irent Infinitive -re Present Participle ant, ante, ants, antes Perfect Participle u, ue, us, ues † ç is treated as c. Mechanical Translation of French 55 If now the French verb stems and their English translations are entered into the machine's dic- tionary together with an instruction that each word fed in for translation is to be identified with the stem in the dictionary which it most closely resembles, we will be supplied with the basic meaning of every verb in the passage to be translated. The tense, however, is still lacking and for this we have to refer to the end- ings. These too need to be included in the dic- tionary, though at the same time they must be kept apart and distinct from the stems; other- wise confusions will arise (e.g. between the singular present indicative of avoir and the endings denoting the singular Past Historic of the 1st conjugation). Nor should they be includ- ed as they appear above for two reasons: Firstly the use of different inflections to indi- cate difference in person and number is on the whole a linguistic redundancy, so that where two identical endings occur we need in fact make only one entry in the dictionary: secondly, inasmuch as the endings are used merely to dis- tinguish tense, there is no necessity to enter the whole ending but only sufficient to enable this distinction to be made (e.g. the -rai part of the ending -evrai in recevoir is enough to indicate that the tense is future). These dis- tinctive endings are 46 in number, i.e. Present e s z t ois oit =stem (is) (it) Future rai ras ra rons rez ront Conditional rais rait rions riez raient Imperfect ais ait ions iez aient Past Historic ai as a mes tes rent is it us ut Perfect Part. é ée é s ées i ie (is) ies ue ues Infinitive r re u (us) Present Part. ant One or two explanatory remarks are necessary here. The "=stem" means that there is no ending (e.g. il vend). The feminine and plural forms of the present participle are omitted because they belong strictly speaking to its ad- jectival function. The forms in parentheses, though characteristic of the tense under which they are marked, are also found in another tense of which it is preferable to regard them as primarily characteristic. This is because the limited capacity of the machine prevents the inclusion of instructions (assuming that they could be formulated) which would enable it to distinguish these formally identical tenses; con- sequently, for the time being one translation must be made to serve for both, and it is na- turally desirable that this translation should be determined by the tense in which this disputed form has the wider occurrence. The ending -is, for example, represents the present indi- cative in finir type verbs, the past historic in finir, vendre and sentir type verbs. Therefore we take it as characteristic primarily of the latter tense; it is just a question again of utility. Clearly, however, the fewer there are of these identical tense forms, the lesser the con- fusion will be. Hence the stem fin- instead of fini- for the finir type verb in the present scheme. If the stem had been fini- the follow- ing difficulties would have arisen: 1. Endings -s and -t 1st, 2nd, 3rd pers. sing. pres. indic. = 1st, 2nd, 3rd pers. sing. past historic 2. Ending -s 1st, 2nd, pers. sing. pres. indic. and past historic = perf. particip. masc. plural 3. No ending: bare stem Sing. perf. particip. masc. = 3rd pers. sing. pres. indic. of vendre type verbs 4. Ending -e and -es Perf. particip. fem. sing. and plural = 1st and 2nd pers. sing. respectively pres. indic. donner type verbs With the stem fin- on the other hand, only the first two of these are encountered. Fin-, there- fore, is the best stem, but only, it must be re- peated, in the present scheme. As the diction- 56 L. Brandwood ary is enlarged and takes in more irregular verbs we may find that there are ambiguities which can be solved by making the stem fini- instead of fin- (e.g. the verb saisir will have a stem identical in form to the 1st and 2nd pers. sing. pres. indic. of savoir, unless the 'i' is added). The important thing to remember is that there can be no one optimum system for all translation programs: their general prin- ciples may very well be the same, but the de- tails must always vary according to the size of the dictionary and the nature of its contents. This applies particularly to stems of verbs, which will continually have to be modified in order to avoid ambiguities (e. g. 1st, 2nd, 3rd pers. sing. pres. indic. of vivre = 1st, 2nd, 3rd pers. sing. past historic of voir, which in its 1st and 2nd persons may again be confused with the stem of the verb viser). Since corre- spondences of this type are far from being un- common, it may prove advisable in a larger dictionary to revert to the conjugational classi- fication each with its separate procedure, in order to reduce the confusion. Returning to our explanation of the table of endings, the reason for using these particular forms to represent the present indicative is to be found in the rule previously mentioned, ac- cording to which the machine identifies any in- coming word or part of a word with the nearest equivalent (by which is meant the nearest shorter equivalent) in its dictionary. There- fore, if we enter, for example, -s in the dic- tionary as a present indicative ending, the machine will identify with it not merely the bare -s of vendre and sentir type verbs, but also the -es of the 1st and the -ons of all con- jugations, simply because there is no other entry in its dictionary with which it can identify them. It is for this reason that the 1st/2nd and 3rd pers. sing. pres. indic. of recevoir type verbs (-ois, -oit) has to be entered in full, since otherwise they would be identified not with the present forms -s and -t, but the past historic -is and -it. The system described above covers not only the regular but also the irregular verbs in the majority of their forms. The only difference is that many of the latter require several stems instead of one; aller, for example, the follow- ing: vais, va-, vont, all-, i Of course, difficulties do arise, just as they do in the re- gular verbs, but their number is relatively small. There are five main types. 1 . One or two verbs such as valoir and vouloir have a 1st and 2nd pers. sing. pres. indic. ending in -x, and one or two such as prendre and asseoir a 3rd pers. sing. in -d. This is solved by adding -x and -d to the list of endings denoting the present tense. 2. There are some verbs with a past historic form which resembles none of the other tenses (except in one or two instances the perfect par- ticiple), and consequently requires a stem pe- culiar to itself. One would think, since this stem is peculiar to a past tense, that its trans- lation in the dictionary could be in the past tense without reference to the ending. For eus, the past historic of avoir, for example, the stem would be eu- and the translation "had" This, however, will not work. The singular with its "present" endings -s and -t comes out all right, but the plural endings -mes, -tes, and -rent are marked in our system as characteristic of the past historic, so that the translation of the plural will be the past tense of "had", which is not at all right. No matter whether we give the stem the equivalent "have" or "had", there will be an anomaly between the translation of singular and plural forms. The solution is to make the stem not eu-, but e The endings will now be found all in the past historic list instead of half in this and half in the present indicative list, and a completely correct translation will be ensured. Besides avoir this solution is applicable to only a few other irregular verbs (e. g. croire, croître, plaire) and not to all those to which it would be desirable to apply it (e.g. être, voir, savoir, mouvoir, pouvoir, boire, faire, lire, prendre, mettre). This is because any stem devised for the past historic of these verbs will be ignored by reason of the rule that incoming words must be identified with the longest similar stem in the dictionary. Let me make this clear with an illustration: for the past historic of voir, i.e. vis, vis, vit, vîmes, vîtes, virent, we can only set up a stem to cover vit (i. e. vi-) vîmes and vîtes (i.e. vî-); whatever we do about the other three they are bound to be con- fused with the stem of either the verb vis-er or the verb vir-er. In cases like this a possible solution is to enter the ambiguous forms in the dictionary as they stand together with an in- struction that if they are found without an end- ing they belong to a different verb and have a different translation than if they are found with an ending. So when vis occurred alone, it would be recognized as belonging to voir, when it occurred with an ending, that is in a longer form, it would be ascribed to viser. The 3rd pers. virent would no longer be ambiguous, if Mechanical Translation of French 57 entered in entirety. Other verbs with past historic forms which cannot be treated in the manner of avoir are tenir and venir (tins and vins) and those which, like finir in the regular verbs, have forms identical in the singular to those of the present tense (e.g. dire, rire, suffire, confire, cir- concire, conclure). The answer for tenir and venir is again to enter the offending forms into the dictionary in their entirety; for dire, rire, etc., as for finir, only an examination by the machine of the context in which they occur will succeed. This, however, lies beyond the scope of our present modest program, which must perforce allow the ambiguity to remain unre- solved. 3. Several verbs (e.g. dire, faire, écrire, conduire, craindre) have a perfect participle identical in form to the 3rd pers. sing. pres. indic. This causes no difficulty in the compound tenses, where the accompanying auxiliary serves to distinguish the participle as such, but when it is used adjectivally, then not only the masc. sing. but also the masc. plur. and the feminine forms will all be interpreted by the machine as present indicatives. A solution by consideration of form is possible only for the unique ending of the masc. plur. -ts (e.g. condui -ts), those of the others being common to other tenses, i.e. the masc. sing.(-t), as we have said, to the 3rd pers.sing. pres. indic., the fem. sing. ( -te) to the 1st and 3rd pers. sing. pres. indic. of 1st conjugation verbs like jeter which double the 't' (i.e. jet-te), and the fem. plur. ( -tes) to the 2nd pers. sing. pres. indic. of these verbs (i.e. jet-tes) or the 2nd pers. plur. past historic of all verbs. An extra stem (i.e. jett-) for the jeter type verbs helps a little, but a complete solution is possible only if the machine can take account of the syntacti- cal context of these adjectival perfect participles. 4. Two of these verbs, faire and dire, present a slightly more complicated problem, since not only is the fem. plur. of their perfect participle identical in form with the 2nd pers. plur. of the pres. indic. (faites and dites), but this form itself has a past historic ending ( -tes). This means that faites and dites have to be entered into the dictionary as they stand, which gets round the second part of the difficulty, leaving the first to be dealt with in the way suggested in no. 3 above. 5. a) The masc. sing. perfect participle of asseoir (assis) and acquérir (acquis) are identi- cal with the 1st and 2nd pers. sing. past histo- ric, whilst the fem. forms of the same have endings which are marked in the above scheme as primarily characteristic of the present indi- cative. b) All perfect participles ending in -u or -i in the masc. sing. will end in -us and -is in the masc. plur. -forms characteristic of the past historic. The solution is again that mentioned in no. 3. In our present program, however, no difficulty is encountered over this similarity of perfect participle and past histo- ric, since we are confining ourselves to indi- cating in translation only whether a verb is past, present, or future. As regards mean- ing no distinction is made between the various forms of the past tense. These, then, are the main difficulties which have been encountered in drawing up a system for translating verbs. Some may have been overlooked, important ones at that I cannot be certain. Of one thing, however, I am quite certain; that is that there is not a single diffi- culty anywhere which is incapable of being sur- mounted. In the method described above, first the meaning of a verb is obtained from the stem, then the tense is discovered from the ending. This is only one way. There is another, as the reader may already have observed for himself in perusing the table of endings, namely by a particular letter in the ending instead of by the whole ending. The most obvious example is the letter 'r', which is sufficient to indicate that the tense is either conditional or future indicative. In place of an ending dictionary, therefore, we can have a series of instructions like this: 1. If the ending ends in -r or -re, then the verb is an infinitive, and "to" should be in- serted before the translation of the stem. The parentheses denote that the "to" will not always be needed to make good English in the transla- tion, e.g., in puis-je sortir? (may I get out), whereas il va partir (he is going to leave) does require it. When larger storage capacities are available, it will be possible to incorporate instructions enabling the machine itself to make this distinction. 2. If the ending ends in -ant, we are dealing with the present participle and "-ing" should be added to the translation of the stem. Again it would be quite easy with a larger storage at one's disposal to include an instruction re- quiring the machine to deduct the mute -e from words ending so, before adding the "-ing". 58 L. Brandwood 3. If the ending contains the letter r and this is not followed by -ent (in which case the tense is the past historic), then the verb in question is either future or conditional. If now the letter i occurs after (but not necessarily immediately after) this r and is not the last letter of the ending, the verb is conditional, and the word "would" must be inserted in front of the stem meaning. Otherwise it is the future tense, and the word to be inserted is "will". (The English idiom of "shall" with the 1st person and the interchange of this and "will" with a change of emphasis could likewise be included in a ma- chine with a larger dictionary space). 4. If the ending begins with e (not followed by a), s (not followed by a or i), t (not followed by anything), x, o , n, 1, d, or if there is no ending (i. e. the stem is bare) and avoir or être does not precede, then the verb is in the present indicative, and the stem meaning is an adequate translation. (The same can be said of the suffixed -s in the 3rd pers. sing. of the English present as of the other refinements mentioned above. ) If there is no ending and avoir or être does precede, the verb is in the perfect tense (see no. 5). The reason why e, s, x, o, d, t indicate the present tense will be clear. The reasons why n and 1 do, and why e should not be followed by a, s, by a or i, or t by any letter are as follows : n- enables certain irregular verbs which double the final consonant of their stem in the 3rd pers. plur. (e. g. venir - viennent) to be comprehended under one stem (i. e. vien-) instead of two. 1- does the same for 1st conjugation verbs in -eler (e.g. atteler - attelle). t is not followed by another letter since -tes is an ending characteristic of the past historic. (Note this means that 1st conjugation verbs in -eter [e. g. jeter - jette] need two stems). e is not followed by a or â because 1st conju- gation verbs in -ger (e. g. manger) retain the e before a and o, so that, but for this precau- tion, the imperfect and past historic of these verbs would be translated as the present tense. s is not followed by a or i, as happens in the imperfect of finir type verbs, if the stem fini- is chosen instead of fin- (i. e. -ssais, -ssait, -ssions, -ssiez, -ssaient. ) (Note: with the present method the stem fini- is preferable to fin- for this reason: if the i were not included in the stem, it would come at the beginning of all forms of the present indi- cative, which would consequently be translated as a past historic. As it is, only the singular forms of the past historic are confused with those of the present indicative. ) 5. Everything not accounted for by the previ- ous instructions is in the past tense, and "-(e)d" should be added to the translation of the stem. (Note: this will result in some odd looking forms in the case of the English strong verbs, but as we have said so often before, it is only the restricted storage capacity of the present machine which forbids the incorporation of these strong forms (e.g. "sang, sung") alongside the regular form ("sing") together with instructions for their correct application.) These are two ways of ascertaining the tense of a verb, which we may call the stem-ending and the eliminating method respectively. The question now arises of which is the better, but this is not easily answered, depending as it does to a great extent on the machine, the size of the dictionary employed, the nature of the text to be translated and the quality of transla- tion required. The eliminating method, for in- stance, must assume that the person receiving the translation will be content to have all past tenses translated simply as such and prepared to use his own common sense and judgement to distinguish where the sense requires a perfect (i.e. has done), imperfect (i.e. was doing), or past historic (i.e. did). If the more precise translation is desired, then there is no doubt that the stem-ending method is to be preferred. As regards the nature of the text, if the tense in which it is written is the past, the stem- ending method will probably be quicker than the eliminating method, which arrives at this tense only at its last stage; if, however, the tense is the present and future, then it may well be that the eliminating method will prove quicker. All this smacks too much of "perhaps" and "maybe", however, as would any remarks which we might hazard on the amount of space which each method would occupy in the ma- chine's dictionary. A definite answer to all such questions will, after all, be forthcoming when the machine starts on its program, since this will be tried with both methods. In the meantime two facts are quite certain: a) the eliminating method is superior to the stem-ending method when it comes to translat- ing perfect participles, because it needs to con- sider only the first letter of the ending, which is the one distinctive of the perfect (i.e. é, i, u) whereas the stem-ending method has to have all the forms (fem. and plur. ) written down for it Mechanical Translation of French 59 to choose from. b) the stem-ending method is more economical for treating the irregular verbs, because the fact that it can ignore the middle part of some such verbs enables it to account for them with one stem less in each case than the eliminating method. The verb asseoir, for example, re- quires three stems in the latter method, i.e. assied- (for the singular of the pres. indic., past historic and conditional), assey- (for the imperf. and plural of the pres. indic.), and ass- (for future, past historic and conditional), but only two in the former method which can ig- nore the -ied- in the sing. pres. indic. and so use ass- for this too. Out of about seventy ir- regular verbs examined, this economizing ap- plied to seventeen, (i. e. acquérir, boullir, tenir, sentir, sortir, partir, servir, asseoir, savoir, pouvoir, écrire, joindre, craindre, mettre, battre, suivre). To these must be added their compounds and any similarly conju- gated verbs, though some of the latter are mis- leading because a verb with a similar stem prevents the reduction of the number of stems. For instance, the only stem that would be needed for dormir would be dor-, but for the existence of a verb dorer; so the stem has to be dorm-, as for the eliminating method, and the sing. forms of the pres. indic. (dors, dort) have to be entered in the dictionary as they stand. That completes our treatment of verbs, and we can pass on now to explain the program for dealing with the other parts of speech. Nouns and adjectives are taken together, since both may have feminine and plural forms, and these can be satisfactorily accounted for by the same stem-ending procedure, thus: Type 1. If the feminine form is created merely by adding extra letters to the masculine form without involving any change in the latter, then the masculine form (so marked) is entered in the stem dictionary, the extra letters de- noting the feminine form ( so marked) in the ending dictionary, (e.g. chien -ne, joli -e). Type 2. If the feminine form is created by first altering the masculine form, then adding to it, (e.g. chameau, chamelle), then the letters which are common to both forms (i.e. chame-) are entered in the stem dictionary, those peculiar to the masculine and feminine in the ending dictionary (i.e. -au, -lle). The gender of a noun which has only one form for both masculine and feminine forms must be de- termined by other means (e.g. by reference to the definite or indefinite article or an adjective, if it is accompanied by any of these). The plural form of nouns and adjectives, both masculine and feminine, is easily recognized as long as there is in fact a distinct plural form to be recognized inasmuch as the ending will have either -x or -s added to it. Apart from irregulars the only class of nouns and ad- jectives which does not comply with this "rule" is that ending in -al, which drops the 1 before adding -ux (e.g. cheval, chevaux). This is re- solved in the same way as the Type 2 feminine form above, that is, by putting the letters common to both singular and plural (in this case cheva- ) in the stem dictionary, the letter 1 which is dropped in the ending dictionary. So we have the following three types of regular nouns and adjectives: Stem Ending Plural Type 1 noun chien (m) -ne (f ) -s adj. joli (m) - e (f) -s Type 2 noun chame- ( -au (m) -x -lle (f) -s adj. bre- -f (m) -s -ve (f) -s Type 3 noun cheva- -1 (m) -ux adj. principa- - 1 (m) -ux -le (f) -s In short, we can identify all plurals which are genuine forms and not identical with the singu- lar by the three suffixes -s, -x, -ux. The only difficulty is in getting the machine to recognize where these suffixes occur after the ending and where directly after the stem. The latter is true of the masculine forms of Type 1 nouns and adjectives and masculine adjectives and all nouns of Type 3; the former of Type 2 nouns and adjectives and Type 3 feminine adjectives. It can be done quite easily if the machine pro- ceeds according to the following instructions, which do not even require it to distinguish be- tween the three suffixes: 1. If when the machine looks in its stem dictionary, a) the incoming noun or adjective can be iden- tified exactly with one of the stems, then it is a Type 1 masculine noun/adjective and can be translated immediately. b) the incoming noun or adjective cannot be identified exactly with a stem, because an ex- tra letter or letters is left over, it cannot be translated immediately; 2. then the machine looks in its ending diction- ary. 60 L. Brandwood a) If the extra letter(s) can now be identified exactly with any ending contained in the diction- ary, then the incoming word is either a femi- nine singular noun/adjective of Type I, a mas- culine or feminine singular noun/adjective of Type II, or a masculine or feminine singular noun/adjective of Type III, and can be trans- lated accordingly (see below); b) the extra letter(s) can be identified with an ending contained in the dictionary, but if there is still an extra letter left over, then the in- coming noun/adjective is the plural of the stem + ending word and its gender is that marked after the ending. The plural is denoted by adding -s to the English translation. c) the extra letter(s) cannot be identified with an ending contained in the dictionary, then the incoming noun/adjective is the plural of the stem word and its gender is that marked after the stem. (Note: the identification of number in adjectives is, of course, irrelevant for their actual translation, but as was noted above, it can be useful for determining the gender and number of nouns with identical forms for the masculine and feminine or the singular and plural. ) This scheme will provide rapid identification of all regular adjectives and nouns, but it has the fault of being somewhat uneconomical as regards storage space, since the ending dic- tionary needs to be of the same size as the stem dictionary, or even larger, most stems having one and often two endings. A possible modification offers itself, however, when we realize that the number of different endings for regular adjectives and nouns is only 15 (5 masculine and 10 feminine) i.e. masculine: -eur, -au, -1, -f, -x; feminine: -euse, -rice, -11e, -ne, -te, -sse, -e, -ve, -se. It would seem feasible, therefore, to have an ending dictionary consisting of only these 14 endings which would be capable of producing a correct translation of all regular French ad- jectives and nouns. And, in fact, it would be, but for English often having a completely dif- ferent feminine form of a noun where French has only a partially different form (e.g. fils, fille son, daughter), or an irregular plural where French has a regular one (e. g. homme, -s man, men). A reasonable solution is to divide all nouns and adjectives into two classes: I. those which have a separate word for mas- culine and feminine, e.g. cheval (horse) - jument (mare), those which are without gender in English, e.g. le village (village), abstract nouns all in fact, which do not require a dis- tinction of gender in translation, (Consequently all adjectives can be included in this class.), II. those which have a feminine or plural form that does require a different translation from the masculine or singular form. The first class, in which the ending is re- quired only for identification not translation, can be dealt with in the manner suggested above, that is by having an ending dictionary for the whole class of only 14 distinct endings (or 28, if we wish to include the plurals and so get immediate recognition of these, rather than provide the machine with instructions). This will retain the speed of identification while at the same time improving the economy. For the second class the unmodified system will be retained, and each noun will be accom- panied by its feminine and plural endings, alongside each of which will be the appropriate translation where different from that of the masculine singular. When the feminine form is merely a lengthened form of the masculine, the two translations will appear with the stem and ending respectively, for example, stem: chien = dog, ending -ne = bitch (dog). When it involves a change in the masculine form, the two translations will accompany the two endings, thus: stem: act- (no translation), endings: -eur = actor, -rice = actress. In closing the section on nouns and adjectives the observation made previously with verbs must be repeated, namely that if the stem- ending method anywhere involves a confusion of two words with different meanings, the stems chosen for these words should be altered, where possible, to create a distinction. Where it is not possible, the offending forms will have to be regarded as irregular and entered in the dictionary in full. The only other part of speech in the program which requires separate instructions for its identification is the pronoun, to be more expli- cit the personal (unstressed) and possessive pronouns. Four difficulties arise, namely those of distinguishing the accusative of the 3rd personal pronoun (le, la, l', les) from the de- finite article, the dative plural of the 3rd per- sonal pronoun (leur) from the identical posses- sive form, and the nominative nous and vous from the accusative/dative, and lastly that of removing the oblique forms of the personal pronoun from their French position before the verb to their normal English position after the verb. The first two difficulties are surmounted by assuming that whenever le, la, l', les are followed by a pronoun or verb, they are forms Mechanical Translation of French 61 of the personal pronoun not the definite article, and similarly that whenever leur is followed by a verb it is the personal pronoun. The last two are also solved simultaneously in the following way: Type 1. the sequence is pronoun l + pronoun 2 + pronoun 3 + verb if nous or vous is the first pronoun, then it is nominative: translate in order - pronoun l + verb + pronoun 2 + pronoun 3 (e.g. nous le leur donnons = we give it them) Type 2. the sequence is pronoun l + pronoun 2 + verb If a) pronoun l is a nominative form (je, tu, il, elle, ils, elles), pronoun 2 is oblique. There- fore translate in order: pronoun l + verb + pronoun 2 (e. g. je le donne = I give it) b) pronoun l is oblique, pronoun 2 is also ob- lique. Translate in order: verb + pronoun l + pronoun 2 (e.g. le berger le lui donne = the shepherd gives it him) c) pronoun l is a form which is not recogni- zable per se as nominative or oblique (i.e. nous or vous), then the verb must be examined to see whether it agrees with nous or vous. Thus: 1) if the verb ends in -ons or -mes, nous- is nominative 2) if the verb ends in -z or -tes, vous is nominative Then translate in order given in a). Otherwise nous and vous are oblique and the translation order is that of b). Type 3. the sequence is - pronoun + verb. The procedure is exactly the same for as Type 2. This, then, is the program which will be used for translating from French mechanically. The objection might well now be raised that it is devoted only to the grammatical side, the parsing of words, and gives no attention to syn- tax. Apart from the fact that an instruction will, as it happens, be included to reverse in translation the order of adjectives following their nouns in French, this objection is quite valid. In defense we must make the plea made so often already that the restrictions imposed by the present limited storage capacity of the computer render any comprehensible scheme for dealing with syntactical problems academic. When the storage capacity is enlarged, as it soon will be, there will be the possibility of making use of such a scheme. Nevertheless, it is probably true to say that, since its word order is more or less identical with that of English, the main difficulty in translating French lies precisely in what we have endeavored to achieve in the program described, namely the successful identification of the various forms of words. This being so, the program will prove capable of producing an adequate translation of such French prose as has no literary pretensions — for example, scientific publications, for which indeed it is primarily intended. Whether this belief is justified or not will be decided by the actual results of the experiment. These will be pub- lished. . longest part of a word common to all forms (inflections) of that word. The stem of donner, for example, is donn-; what remains of each inflection of donner. successful identification of the various forms of words. This being so, the program will prove capable of producing an adequate translation of such French prose